DBNet: Real-time Scene Text Detection with Differentiable Binarization

Segmentation detectors predict, for every pixel, the probability that it belongs to text. Turning that probability map into boxes needs a threshold, and before DBNet (Liao et al., AAAI 2020) the threshold was a fixed number chosen by hand. DB adds a second output, a threshold map, and combines the two with a steep sigmoid so that binarization becomes a differentiable operation the network can learn. At inference the threshold branch is dropped; a plain threshold on the probability map, one contour pass and a polygon “unclip” step produce the final shapes.

Why every OCR toolkit picked it

• Accuracy and speed together. The abstract reports an F-measure of 0.828 at 62 FPS on MSRA-TD500 with a ResNet-18 backbone.
• Arbitrary shapes. Because the output is a mask, curved and rotated text come out as polygons with no special cases.
• Tiny post-processing. No affinity maps, no character grouping, no NMS.

PaddleOCR’s PP-OCR detectors, MMOCR’s default detector, docTR’s db_resnet50 and OpenCV’s TextDetectionModel_DB are all DB-family models. The detector on this site runs PP-OCRv6 and PP-OCRv5 detection models, which are DB-style networks with lighter backbones, exported to ONNX.

DB++ (2022)

Real-Time Scene Text Detection with Differentiable Binarization and Adaptive Scale Fusion (Liao et al., TPAMI 2022, arXiv:2202.10304) adds an Adaptive Scale Fusion module that weights multi-scale features before segmentation, improving robustness to text size. The DB idea is unchanged.

Knobs you will meet in code

The OpenCV guide shows these in a runnable script.

Citation

@inproceedings{liao2020db,
  title     = {Real-time Scene Text Detection with Differentiable Binarization},
  author    = {Liao, Minghui and Wan, Zhaoyi and Yao, Cong and Chen, Kai and Bai, Xiang},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year      = {2020}
}